Comparative population genomics of maize domestication and improvement

Abstract

Domestication and plant breeding are ongoing, 10,000-year evolutionary experiments that have radically altered wild species to meet human needs. Maize has undergone a particularly striking transformation. Researchers have sought for decades to identify the genes underlying maize evolution^1,2, but these efforts have been limited in scope. Here, we report a comprehensive assessment of the evolution of modern maize based on the genome-wide resequencing of 75 wild, landrace, and improved maize lines³. We find evidence of recovery of diversity post-domestication, likely introgression from wild relatives, and evidence for stronger selection during domestication than improvement. We identify a number of genes with stronger signals of selection than those previously shown to underlie major morphological changes4,5. Finally, through transcriptome-wide analysis of gene expression, we find evidence consistent with removal of cis-acting variation during maize domestication and improvement and suggestive of modern breeding having increased dominance in expression while targeting highly expressed genes.

Archaeological⁶ and genetic^7,8 evidence indicate that maize (Zea mays ssp. mays) was domesticated approximately 10,000 B.P. in the Balsas River Basin of southwestern Mexico. Domestication involved a radical phenotypic transformation from the wild progenitor, Zea mays ssp. parviglumis (hereafter parviglumis; Fig. 1), resulting in an unbranched plant with seed attached to a cob, making maize entirely dependent on humans for propagation. Subsequent to domestication, maize has been subject to intensive improvement efforts, culminating in the development of hybrid maize lines highly adapted to modern agricultural practices. We present a population genomic analysis of maize evolution based on resequencing of 75 genomes of maize and its wild relatives (Fig. 1, Supplementary Fig. 1, and Supplementary Table 1). We generated 781 gigabases of sequence from 35 improved maize lines, 23 traditional landraces, and 17 wild relatives (14 parviglumis; two Zea mays ssp. mexicana, hereafter mexicana; one Tripsacum dactyloides var. meridionale) using short-read technology, sequencing each line to an average depth of more than 5× (Supplementary Table 1; ref. 3). Reads were mapped to the maize reference genome (release 4a.53) and analyses are based on a final set of 21,141,953 high quality single nucleotide polymorphisms (SNPs).

Figure 1. Neighbor-joining tree and changing morphology of domesticated maize and its wild relatives.

Taxa in the neighbor-joining tree (right) are represented by different colors: parviglumis (green), landraces (red), improved lines (blue), mexicana (yellow), and Tripsacum (brown). Morphological changes (left) are shown for female inflorescences and plant architecture during domestication and improvement.

Maize landraces retain more nucleotide diversity (83%; Fig. 2a) and show lower genetic differentiation from their wild progenitor (F_ST=0.11) than other crop species^9,10. This is likely due to large census size and an outcrossing mating system in maize landraces. Linkage disequilibrium has increased dramatically as a result of domestication, with genome-wide estimates of the population recombination rate ρ in landraces estimated to be 25% of the rate in parviglumis and average haplotype length increasing from 22 to 30kb (Supplementary Fig. 2). These results are consistent with the effects of a domestication bottleneck, but an excess of rare SNPs (Supplementary Table 2 and Supplementary Fig. 3) suggests that variation has begun to recover across most of the genome. Gene-rich regions, however, exhibit fewer SNPs unique to landraces (t-test, p<0.001) and no excess of rare SNPs (Supplementary Tables 2 and 3), a difference likely due to the effects of background selection against deleterious mutations slowing the post-domestication recovery of variation at linked sites. Modern breeding appears to have had negligible effects on genome-wide diversity or mean haplotype lengths in our broad sample of modern lines (Fig. 2a, Supplementary Fig. 2, Supplementary Table 2). While our estimates of nucleotide diversity in improved lines may be inflated by the diverse inbreds chosen, the relationships between inbred and landrace lines (Fig. 1) suggest a weaker genome-wide bottleneck during improvement. Finally, comparison of maize landraces to our two mexicana genomes identifies several extended regions of high genetic similarity (Supplementary Fig. 4), consistent with previous observations of admixture between these taxa^8,11 and suggestive of the possibility that mexicana may have contributed alleles important for maize evolution.

Figure 2. Genome-wide analysis of nucleotide diversity and selection.

(a) LOWESS curves of nucleotide diversity (π) along chromosome 1 in parviglumis (green), landraces (red), and improved lines (blue). Genome-wide likelihood values (XP-CLR) for selection during domestication (b) and improvement (c) with chromosome number indicated along the x-axis. Distributions of feature size (d) and gene counts within features (e) in domestication and improvement scans.

To identify regions of the genome most affected by selection during maize evolution, we used a likelihood method (XP-CLR; ref. 12) to scan for extreme allele frequency differentiation over extended linked regions (Fig. 2b, c). Adjacent windows of high XP-CLR were grouped into “features,” each likely representing the effect of a single selective sweep. Features in multiple centromeres show high XP-CLR values (Fig. 2b, c, and Supplementary Fig. 5). Combined with evidence for change in abundance of centromeric retroelements (Supplementary Table 4 and Supplementary Fig. 6; ref. 3), this finding suggests rapid centromere evolution. However, because centromeres harbor few genes and our genetic map may underestimate the extended LD in these regions, we masked centromeres from further analysis. We also masked a newly discovered ~50-Mb inversion polymorphism on chromosome 1 (Supplementary Fig. 7; further characterized in submitted publication by Z. Fang, T.P., A.L. Weber, R.K. Dawe, J.C.G., J. Sánchez-González, C. Ross-Ibarra, J.D., P.L.M., J.R.I.).

We focused analyses on the 484 domestication and 695 improvement features in the highest 10% of XP-CLR values (Fig. 2b, c). Domestication features contain an average of 3.4 genes and have a mean size of 322 kb (Fig. 2d, e), cover approximately 7.6% of the maize genome, and show multiple signatures of selection, including elevated differentiation, low nucleotide diversity, and an excess of high-frequency derived SNPs (Supplementary Table 5 and Supplementary Fig. 8). We estimate the mean strength of selection in these features as s = 0.015, which is within the range of estimates based on archaeological data from other domesticates¹³ and more than an order of magnitude higher than the mean value of 0.0011 across the rest of the genome.

While selection during maize improvement can be strong^14,15, XP-CLR values and estimated selection coefficients (mean s = 0.003) from our improvement scan were substantially lower than observed for domestication (Fig. 2b, c). Consistent with this finding, improvement features have smaller average size (Fig. 2d) and contain fewer genes (Fig. 2e) than domestication features. One explanation for these results may be that our diverse tropical and temperate lines derive from distinct landrace founders (Fig. 1) and have been subject to different selective pressures¹⁶. Indeed, independent scans of temperate and tropical lines find stronger evidence of selection and little overlap of selected features (Supplementary Fig. 9). However, previous estimates of effect size for loci involved in domestication and improvement traits provide some independent evidence of stronger selection during domestication¹⁷. Twenty-three percent (107) of domestication features show additional evidence of selection during improvement, indicating that a subset of domestication loci may contribute to phenotypes of continued agronomic importance.

Individual features likely result from a single selective event, and we assigned the gene closest to the 10-kb window with the maximum XP-CLR score in each feature as the most likely candidate (Supplementary Tables 6 and 7). Our domestication and improvement candidate lists, each including 1–2% of the maize filtered gene set (FGS), represent our best estimate of the direct targets of selection within features, but linked genes have also been affected by selection, limiting the diversity available for modern improvement for many of the 3,040 genes found within features. Thus, while our candidates are of most interest for understanding genes directly related to maize evolution, breeding programs would likely benefit from efforts to incorporate diversity from exotic germplasm in these genomic regions.

A sizeable fraction of our domestication and improvement features contain no annotated sequence (6% and 11%), a result that could implicate regulatory variants in the process of maize evolution. However, the majority of our features contain genes in the high-confidence FGS, and should prove useful both in dissecting existing QTLs and identifying novel candidate genes. For example, the domestication candidate GRMZM2G448355, an ortholog of the rice gene OsMADS56 which delays flowering under long-day conditions (Fig. 3 a–c), is found within a flowering-time QTL on chromosome 9¹⁸ and two improvement candidates implicated in nitrogen metabolism, GRMZM2G036464 (glutamine synthetase) and GRMZM2G428027 (nitrate reductase), both reside in a QTL for multiple traits including thousand kernel weight and nitrogen mobilization¹⁹. Only a fraction of our novel candidate genes are functionally characterized in maize; one example is the domestication candidate abph1 (GRMZM2G035688) that is known to affect phyllotaxy²⁰. However, function can often be inferred from orthology; the domestication candidate GRMZM2G010290 has no known function, but shows close sequence identity to Arabidopsis DAG1 and DAG2 proteins that affect seed germination²¹. Two improvement candidates, gibberellin 2-oxidase (GRMZM2G152354) and gibberellin 3-oxidase (GRMZM2G036340, dwarf1) are found in the plant growth hormone gibberellin biosynthesis pathway (Fig. 3 d–f) upstream and downstream of the “green revolution gene,” gibberellin 20-oxidase²². Other notable improvement candidates include GRMZM2G082468, a homolog of Arabidopsis farnesyltransferase that has been engineered as a drought tolerance transgene in canola²³, and GRMZM2G087612, whose Arabidopsis ortholog, SDP1, initiates storage oil breakdown in seed²⁴.

Figure 3. Domestication (a–c) and improvement (d–f) candidate genes in relation to two pathways in rice.

Zea mays genes are on a grey background (candidates) or boxed (in selected regions) above their rice orthologs. (a) The flowering-time pathway²⁹, including GRMZM2G448355 and zagl1. (b) Seedling expression pattern of GRMZM2G448355 in parviglumis and maize inbreds. (c) XP-CLR and relative diversity near GRMZM2G448355; gene orientation is indicated with arrows. (d) The gibberellin biosynthesis pathway³⁰. (e) and (f) XP-CLR near the improvement candidates GRMZM2G152354 and dwarf1. The high-yielding rice variety IR8 has a mutation in GA20ox²².

To further characterize the genomic impact of domestication, we used long-oligo array hybridization to survey expression of 18,242 genes in the FGS in seedling tissue of a subset of 25 improved maize and seven parviglumis lines (Supplementary Table 1). Compared to non-candidates, our domestication candidates show greater absolute change in expression between parviglumis and maize (29% versus 22% in non-candidates, p=0.004, Supplementary Fig. 10), upregulation in maize relative to parviglumis (11.4% of candidates upregulated versus 6.5% of non-candidates, p=0.001, Supplementary Fig. 10) and a 10% lower coefficient of variation in expression among maize lines (p=0.006). Reduced variation in expression is observed throughout candidate features, suggesting removal of cis-variation at sites linked to the target of selection. Improvement candidates also show decreased variation in expression in maize relative to parviglumis (8% reduction in CoV, p=0.019), but do not show a significant change in the magnitude of expression. While the reduction in variation in expression could be due to selection on linked sites, the directional change seen in domestication candidates suggests the action of selection on cis-acting regulation.

Although changes in expression during domestication are unlikely to be limited to seedling tissue, our domestication candidates show no tissue-specific patterns of expression (Supplementary Fig. 11a; ref. 25). Improvement candidates also show no tissue specificity, but are more highly expressed than non-candidates in all but one of the tissue groups evaluated (p=0.025–0.044; Supplementary Fig. 11b). Because improvement candidates show no significant change in expression between teosinte and modern inbreds, this latter result suggests that modern maize improvement may have targeted loci that were already highly expressed. Comparison to the full FGS finds no evidence for an overall bias towards constitutive expression in our candidates, in contrast to previous resequencing scans^2,26 (Supplemental Table 8).

Finally, we took advantage of expression data from crosses between inbred lines to evaluate levels of dominance in our candidate genes²⁷. Domestication candidates show elevated dominance (p=0.001), but no significant difference in dominance (t-test, p=0.74) between crosses from the same or different genetic (heterotic) groups, a result that can be explained simply by the loss of additive cis-regulatory variation. Improvement candidates, in contrast, show higher dominance of expression than non-candidates (p=0.007) mostly due to higher dominance in crosses between heterotic groups (p=0.001), likely reflecting the important role that complementation between heterotic groups has played in maize improvement.

Our comparative genomic analysis of wild, landrace, and modern maize sheds light on the complexities of crop evolution and offers guidance to modern breeding. George Beadle²⁸ and others^4,5 have shown that a few genes (e.g., tb1 and tga1) radically altered some aspects of morphology during domestication. The majority of domestication features we identify show stronger evidence of selection than these canonical domestication genes, implying domestication targeted hundreds of genes of diverse biological function that likely involved unstudied aspects of phenotype. The loss of diversity at sites linked to selection and the observed enrichment of improvement candidates for highly expressed genes suggests that modern breeding has mostly worked with the “low-hanging fruit” of the genome and that much could be gained by focusing breeding efforts on the effective incorporation of diversity at other loci.

Methods

Methods and any associated references are available in the online version of the paper at http://www.nature.com/naturegenetics/.

Methods

Sequencing and read mapping

As part of the maize Hapmap II project, sequence was generated and paired-end libraries were prepared in a subset of 75 of the 103 lines described in Chia et al.³ Sampled accessions include 35 improved maize lines, 23 traditional landraces, and 17 wild relatives but exclude 28 tropical inbreds from the International Maize and Wheat Improvement Center (CIMMYT) sampled by Chia et al.³ Sequence depth for each line is detailed in Supplementary Table 1. Lines were sequenced to a mean depth of 5.05X (median 4.58X) with mean coverage in analyzed regions of 3.74X (median 3.70X). Paired-end sequence was aligned to the AGPv1 B73 reference genome³¹, and SNPs called following algorithms described in Chia et al.³ We further filtered SNPs by retaining only non-singleton, biallelic SNPs with ≤ 50% missing data across all three groups of interest (parviglumis, landraces, and improved lines) resulting in a final dataset of 21,141,953 SNPs.

Error rate estimation

Per-nucleotide error rates for the HapMap II data set were estimated to be ~0.1% when compared to Sanger-sequenced BACs, on par with ~0.12% error across loci genome-wide for Sanger resequencing data³. To estimate the genotypic error rate in our subset of typed SNPs, SNP calls from the maize Hapmap II and maize Illumina Infinium 55K chip datasets were compared for a subset of 66 lines (SNP data for the 55K lines are available at www.panzea.org). The Infinium 55K data set is comprised of 51,584 SNPs, of which 37,279 were in common with the subset of Hapmap II SNPs used in our study. Out of 2.2 million comparable genotypes, 2.45% differed between the SNP calls from the two datasets. Most (88%) of the genotypic discrepancies consisted of a heterozygous versus homozygous call; hence, if we assume no errors in the Infinium 55K data set, the mean and median homozygous error rates in our data set per line (0.31% and 0.20% respectively) were much lower than the overall genotypic error rates and the actual perallele error rate was ~1%. Most (59.1%) SNPs had a genotypic error rate of 2% or less, 30.2% of SNPs had zero errors, and only 1.5% of SNPs had an error rate higher than 10%. Both the genotypic and homozygous error rates followed a unimodal distribution with a mode of zero, and there was no obvious second class of SNPs with poor performance with respect to error rates. B73 had the lowest genotypic error rate (0.87%) and the third lowest homozygous error rate (0.12%); this suggests that most of the genotyping errors in the Hapmap II data resulted from alignment issues rather than raw sequencing errors. The four teosinte lines with the highest homozygous error rates (TIL01, TIL07, TIL08, and TIL09) were represented in the two datasets by individuals of different selfing generations (e.g., S5 vs. S8 for TIL08).

Genome scan for selection

We performed a genome scan using the composite likelihood approach (XP-CLR) of Chen et al.¹², modified to incorporate missing data (code available on request). Evidence for selection across the genome during domestication and improvement was evaluated in two contrasts: landraces vs. parviglumis for domestication and improved lines versus landraces for improvement. Our scan used a 0.05-cM sliding window, stepping every 100 bp across the genome. Individual SNPs were assigned a position along the genetic map³² by assuming uniform recombination between mapped markers. To ensure comparability of the composite likelihood score in each window, we fixed the number of SNPs assayed in each window at 50. Following Chen et al.¹², we down-weighted pairs of SNPs in high LD (R² > 0.70) to minimize the effect of dependence on the composite likelihood score. Final estimates were tabulated in non-overlapping 10-kb windows across the genome, assigning each 10-kb window the mean likelihood score (XP-CLR) and selection coefficient (s) estimated by the method of Chen et al.¹²

To partially account for the non-independence of XP-CLR scores along the physical map, we grouped regions into putatively selected “features”. Features were defined as groups of 10-kb windows with XP-CLR values above the genome-wide 80^th percentile uninterrupted by more than one window below this threshold. Features falling within 0.05 cM of functional centromeres³³ and an inversion on chromosome 1 (Supplementary Note) were masked from subsequent analyses. Our analyses of regions selected during domestication and improvement focused on features in the highest 10^th percentile of mean feature-wise XP-CLR. However, we applied a more stringent criterion for identifying candidate genes, drawing only from features in which the observed reduction in nucleotide diversity during domestication or improvement was lower than the median observed from 1000 random windows of similar width and nucleotide diversity. We assigned the maize FGS gene closest to the window with the maximum XP-CLR value as the most likely candidate.

Population genetic analyses

Individual SNPs for each gene were classified as noncoding, synonymous coding, or nonsynonymous coding based on annotations of the first transcript in the FGS. Standard population genetic summary statistics (π, ρ, F_ST, Tajima’s D, Fay and Wu’s H′) were calculated for non-overlapping, 10-kb windows across the genome and separately for individual genes in the FGS using a combination of custom scripts, programs written using the libsequence C++ library³⁴, and SAMtools³⁵. In addition to statistics within groups, we calculated a weighted F_ST³⁶ between groups and net pair-wise divergence between parviglumis and the outgroup Tripsacum. We tested for outliers of our summary statistics in candidate regions by comparing average values to a distribution calculated for randomly sampled, non-overlapping, genomic regions of identical width. Site frequency spectra were rescaled to address missing data and differing sample sizes (Supplementary Note).

To estimate mean haplotype lengths in each group, we used a custom perl script to choose 1 million starting points uniformly across the genome. At each point we chose two random lines from within a group (parviglumis, landraces, improved lines). We compared the two lines’ genotypes at the focal point, extending outward in both directions until we found different genotypes. Missing data and heterozygous SNPs were ignored.

Historical recombination rates (ρ=4Ne) were estimated using the composite likelihood approach of Hudson³⁷. For estimating recombination, we treated the data as haploid, coding all heterozygous sites as missing data. We subsequently removed all SNPs with minor allele counts of ≤ 2. For windows with ≥ 10 remaining SNPs, values of ρ per bp were estimated across a grid of values from 10⁻⁴ to 0.2, assuming no homologous gene conversion

Expression analyses

Three separate datasets were used to assess patterns of gene expression in maize and parviglumis transcriptomes. First, using a custom long oligonucleotide microarray³⁸ designed by NimbleGen (GPL10846, Roche NimbleGen, Madison, WI), we characterized variation in gene expression in a subset of our 25 improved maize and 7 parviglumis inbred lines (Supplementary Table 1). Multiple replications of the maize inbred lines B73 and Mo17 were included to assess consistency. Plants were grown and seedling leaf tissue was harvested 8 days after germination. RNAs were isolated using the commercial TRIzol (Invitrogen, Carlsbad, CA; cat# 15596026) from above-ground tissue and purified by Lithium Chloride treatment followed by 3 M sodium acetate (0.1 vol) and 95% ethanol (2.5 vol) precipitation. Purified RNAs (10 ug) were reverse transcribed and labeled according to the array manufacturer protocol. Per sample, ~20 µg of Cy3- or Cy5-labeled RNAs were hybridized for 16–20 hours at 42 °C using the NimbleGen Hybridization System. Post hybridization, slides were washed (NimbleGen Wash Buffer Kit) and dried for two minutes by centrifugation. Slides were immediately scanned using the GenePix 4000B Scanner (Molecular Devices, Sunnyvale, CA) according to the array manufacturer protocol.

Array images and data were processed using NimbleScan software. Briefly, images from each slide were separated into 12 subarrays and aligned to a grid to extract signal intensity for each feature on the array. Experimental integrity was verified by evaluation of the signal intensities of the sample tracking control features for each subarray. In addition, metrics reports were produced for each array to describe the signal uniformity across the array and the intensity of known empty features, random probes and experimental probes. Signal-to-noise ratios were estimated by dividing the average signal intensity of experimental gene probes by that of the control probes. Only slides with a signal-to-noise ratio ≤ 2 were retained. NimbleScan was used to generate RMA-normalized³⁹ gene expression values from the spatially-corrected probe signal intensities on a per probe and per gene basis. Normalized gene expression values across multiple replications (technical or biological) of the same genotype were averaged, when possible. Comparisons of the distributions of signal intensity for control and experimental probes (Supplementary Fig. 12) were used to determine a reasonable signal threshold for positive expression across all slides. Genes with average probe log2 signals of > 10 in at least three arrays were retained as expressed (N=19,792 expressed genes).

Nucleotide polymorphism may contribute to differences in hybridization between transcripts from divergent genotypes, though previous results suggest as many as 4–5 SNPs are needed to strongly affect probe hybridization⁴⁰. To minimize the impact of polymorphism on hybridization, we further filtered the probeset based on a previous comparative genomic hybridization dataset developed using many of the same genotypes and the same array platform³⁸. Probes that exhibited substantially reduced CGH values for at least three genotypes (26,937 probes) were removed resulting in a set of 46,167 probes that detected expression of 18,242 genes with one to four probes per gene. This subset of the data was used for all subsequent analyses. Finally, a Spearman’s rank correlation showed a weakly positive correlation between parviglumis expression and F_ST parviglumis/maize (rho = 0.043) providing no evidence for hybridization bias.

Our second data set comprised expression estimates in 60 different tissues of the inbred line B73 from a NimbleGen array of 23,740 genes in the FGS²⁵. These data were curated and categorized following conventions described in Sekhon et al.²⁵.

The third dataset consisted of expression from 5 improved lines and their F1 hybrids characterized using an Affymetrix GeneChip® Maize Genome Array (GMGA)²⁷. Probesets with fewer than two biological replicates were dropped, and remaining probesets were annotated by comparing the AGPv1 physical map positions of array probes to the FGS⁴¹. Probesets were included in the analysis only if they mapped to a single gene. When multiple probesets mapped to a single gene, expression data from all probesets were averaged for further analysis.

Significant differences between candidate and non-candidate expression values were determined by bootstrap resampling of log2-transformed, RMA-normalized data from non-candidate genes. The validity of a bootstrap approach was assessed by plotting the mean of each bootstrap sample against its variance in order to confirm calculated test statistics were pivots⁴². For tissue-specific expression, bootstrap significance values were adjusted for multiple tests with a Benjamini-Hochberg False Discovery Rate correction at the 0.05-level. Dominance was assessed as:

$$(\text{hybrid expression signal}-\text{midparent expression signal})/(\text{range of parental expression}/2)$$

Finally, in order to ensure that the relatively low coefficient of variation in expression observed in candidates did not result in an inflation of estimates of dominance, an analysis of covariance was conducted with the coefficient of variation included as a covariate.